Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Optimal Trade Execution: A Mean–Quadratic-Variation Approach ∗1
P.A. Forsyth † J.S. Kennedy ‡ S. T. Tse § H. Windcliff ¶2
September 25, 20113
Abstract4
We propose the use of a mean–quadratic-variation criteria to determine an optimal trading strategy5
in the presence of price impact. We derive the Hamilton Jacobi Bellman (HJB) Partial Differential6
Equation (PDE) for the optimal strategy, assuming the underlying asset follows Geometric Brownian7
Motion (GBM) or Arithmetic Brownian Motion (ABM). The exact solution of the ABM formulation is8
in fact identical to the static (price-independent) approximate solution for the mean-variance objective9
function in Almgren and Chriss (2001). The optimal trading strategy in the GBM case is in general a10
function of the asset price. The static strategy determined in the ABM formulation turns out to be an11
excellent approximation for the GBM case, even when volatility is large.12
Keywords: optimal trading, mean–quadratic-variation, HJB equation13
JEL Classification: C63, G1114
1 Introduction15
A typical problem faced by an investment bank arises when buying or selling a large block of shares. If the16
trade is executed rapidly, then this can be expected to cause a significant price impact. For example, in the17
case of selling, this price impact will lower the average price received per share compared to the pretrade18
price. An obvious strategy is to break up the trade into a set of smaller blocks. This will lower the price19
impact, but now the trading takes place over a longer time horizon. Consequently, the seller is exposed to20
risk due to the stochastic movement of the stock price, relative to the pretrade price.21
Algorithmic trading strategies attempt to determine a trading schedule which optimizes a given objective22
function. One of the early papers on this topic (Bertsimas and Lo, 1998) considered the best trading strategy23
which minimizes the cost of trading over a fixed time. More recently, this problem has been posed in terms24
of a mean-variance tradeoff in continuous time (Almgren and Chriss, 2001; Almgren, 2003; Almgren et al.,25
2004; Engle and Ferstenberg, 2007; Lorenz, 2008; Lorenz and Almgren, 2011). Another possibility is to26
maximize an exponential or power law utility function (He and Mamaysky, 2005; Vath et al., 2007; Schied27
and Schoneborn, 2008). However, the mean-variance tradeoff has a simple intuitive interpretation, and is28
probably preferred by practitioners.29
∗This work was supported by the Natural Sciences and Engineering Research Council of Canada, and by a Morgan StanleyEquity Market Microstructure Research Grant. The views expressed herein are solely those of the authors, and not those ofany other person or entity, including Morgan Stanley. Morgan Stanley is not responsible for any errors or omissions. Nothingin this article should be construed as a recommendation by Morgan Stanley to buy or sell any security of any kind.†David R. Cheriton School of Computer Science, University of Waterloo, Waterloo ON, Canada N2L 3G1
[email protected]‡Morgan Stanley, 1585 Broadway, New York, NY 10036, [email protected]§David R. Cheriton School of Computer Science, University of Waterloo, Waterloo ON, Canada N2L 3G1
[email protected]¶Morgan Stanley, 1585 Broadway, New York, NY 10036, [email protected]
1
While single period mean-variance optimization is straightforward to understand, multi-period mean-30
variance optimization is more complex. In particular, there are several formulations of multi-period mean-31
variance optimization in the literature. The formulation that aguably aligns best with performance measures32
used in practice is the mean-variance criteria as seen at the start of the trade execution. This formulation is33
termed the pre-commitment mean-variance formulation in the literature. (Basak and Chabakauri, 2010)34
One subtlety of the pre-commitment mean-variance formulation is that it gives rise to optimal strategies35
that are time-inconsistent. Consequently, the dynamic programming principle cannot be directly applied to36
solve for the optimal strategies. In view of this difficulty, various approximations to the pre-commitment37
mean-variance formulation have been proposed. Our work is a step toward understanding how such approx-38
imate strategies compare with the the truly optimal strategies.39
In particular, we offer a new perspective on the Almgren and Chriss (2001) approximation. The approach40
in Almgren and Chriss (2001) restricts attention to static strategies (i.e. do not use any information regarding41
the stock price evolution after the start of trading). For static strategies, it turns out that variance, which42
is the original risk measure intended to be minimized, becomes equal to the expected value of quadratic43
variation. Nevertheless, we emphasize that this equality, by itself, does not imply that the static solution44
in Almgren and Chriss (2001) is optimal for the quadratic variation risk measure. The subtle point here is45
that Almgren and Chriss (2001) does not study how dynamic strategies perform in terms of the quadratic46
variation risk measure.47
We thus suggest directly formulating the optimal trade execution problem in terms of a mean quadratic48
variation objective function. We discuss the conceptual simplicity of the mean quadratic variation formu-49
lation, and put forward the idea that quadratic variation is a sensible risk measure. We also discuss the50
properties of the optimal strategies under this risk measure. Our earlier paper (Forsyth, 2011) focuses on51
solving for the truly optimal solutions for the pre-commitment mean variance formulation. In addition, (Tse52
et al., 2011) compares the optimal strategies determined by the pre-commitment mean variance formulation53
and the mean quadratic variation formulation.54
In summary:55
• The mean quadratic variation formulation is conceptually simpler to understand than the pre-commitment56
mean variance formulation. In particular the mean quadratic variation objective leads to optimal57
strategies that are clearly time-consistent and can be easily obtained using the dynamic program-58
ming principle. In contrast, the pre-commitment mean variance formulation has optimal strategies59
that are time-inconsistent, and these strategies are considered unnatural by some authors (Basak and60
Chabakauri, 2010).61
• As a risk measure, quadratic variation has the property that it takes into account the trajectory of the62
liquidation profile, as opposed to variance which measures only the end result with no concern of how63
liquidation proceeds during the whole trading horizon. Alternatively, minimizing quadratic variation64
can also be seen as minimizing the volatility of the portfolio value. Note that one purpose of converting65
shares to cash is to make the portfolio value process less volatile.66
We also mention that Gatheral and Schied (2010) suggest time averaged value at risk as another risk measure67
that also leads to time-consistent strategies.68
The main contributions of this article are69
• We formulate the mean quadratic variation problem in which quadratic variation, rather than variance,70
is used as the risk measure. We argue that quadratic variation can be regarded as a reasonable risk71
measure in optimal trade execution. We derive the Hamilton Jacobi Bellman (HJB) partial differential72
equations (PDE) and provide numerical methods to solve for both the optimal strategies and the73
efficient frontier, with arbitrary constraints on the strategy.74
• We formulate the optimal trade execution problem assuming that the asset price dynamic follows either75
Geometric Brownian Motion (GBM) or Arithmetic Brownian Motion (ABM). We believe that GBM is76
a superior model since it can be used for both long and short trading horizons, and avoids the fallacy77
2
of negative price scenarios that can appear in ABM. Nevertheless, we also study the ABM case since78
this allows us to compare our results with those in Almgren and Chriss (2001).79
• We provide a proof that the classic static solution in Almgren and Chriss (2001) is optimal under the80
mean quadratic variation formulation, even when optimization is over the class of dynamic strategies81
(in the ABM case). The static solution in Almgren and Chriss (2001) is originally obtained as an82
approximate solution to the pre-commitment mean variance problem by restricting attention to static83
strategies. As such, the static strategy is strictly suboptimal in the pre-commitment mean variance84
formulation, since the optimal strategies are dynamic (Forsyth, 2011; Lorenz and Almgren, 2006;85
Lorenz, 2008). Therefore, our proof shows that the classic static solution in Almgren and Chriss86
(2001) is actually an exact optimal solution to the mean quadratic variation problem assuming ABM.87
• We show that in the mean quadratic variation formulation, optimal strategies in the GBM case are88
qualitatively different from those in the ABM case. More specifically, optimal strategies are dynamic89
in the GBM case, whereas optimal strategies are static in the ABM case. This contrasts with the pre-90
commitment mean variance formulation in which optimal strategies are dynamic in both the GBM91
case and the ABM case.92
• Our numerical results show that if we use the optimal static strategies from the ABM case as approx-93
imate solutions for the GBM case, this results in an efficient frontier that is very close to the true94
efficient frontier. While the accuracy of this ABM approximation is obvious when the ABM and the95
GBM dynamics are close, it is surprising that the accuracy of the ABM approximation is excellent even96
when volatility is very large. We explain this in detail and note that the accuracy of this approximation97
does not hold in the pre-commitment mean variance formulation.98
2 Optimal Execution99
Let100
P = B +AS = Portfolio
S = Price of the underlying risky asset,
B = Balance of risk free bank account,
A = Number of shares of underlying asset.
The optimal execution problem over t ∈ [0, T ] has the initial condition101
S(0) = sinit, B(0) = 0, A(0) = αinit. (2.1)
If αinit > 0, the trader is liquidating a long position (selling). If αinit < 0, the trader is liquidating a short102
position (buying). In this article, for definiteness, we consider the selling case. At t = T ,103
S = S(T ), B = B(T ), A(T ) = 0, (2.2)
where B(T ) is the cash generated by selling shares and investing in the risk free bank account B, with a final104
liquidation at t = T− to ensure that A(T ) = 0. The objective of optimal execution is to maximize B(T ),105
while at the same time minimizing a certain risk measure.106
In this paper, we consider Markovian trading strategies v(·) that specify a trading rate v as a function of107
the current state, i.e. v(·) : (S(t), A(t), t) 7→ v = v(S(t), A(t), t) . Note that in using the shorthand notations108
v(·) for the mapping, and v for the value v = v(S(t), A(t), t), the dependence of v on the current state is109
implicit.110
By definition,111
dA(t) = v dt. (2.3)
3
We assume that due to temporary price impact, selling shares at the rate v at the market price S(t) gives112
an execution price Sexec(v, t) ≤ S(t). It follows that113
dB(t) =(rB(t)− vSexec(v, t)
)dt (2.4)
where r is the risk free rate.114
We suppose that the market price of the risky asset S follows a Geometric Brownian Motion (GBM) or115
arithmetic Brownian Motion (ABM), where the drift term is modified due to the permanent price impact of116
trading.117
In the GBM model, we assume118
dS(t) =(µ+ g(v)
)S(t) dt+ σS(t) dW(t),
µ = drift rate,
g(v) is the permanent price impact function,
σ = volatility.
W(t) is a Wiener process under the real world measure. (2.5)
In the ABM model, we assume119
dS(t) =(µ+ g(v)
)S(0) dt+ σS(0) dW(t) . (2.6)
3 Price Impact120
In this section we specify the permanent and temporary price impact functions used in this paper. We refer121
the reader to Almgren et al. (2004) for a discussion of the rationale behind these permanent and temporary122
price impact models.123
In both the GBM case and the ABM case, we use the following form for the permanent price impact124
g(v) = κpv,
κp is the permanent price impact factor . (3.1)
We take κp to be a constant. This form of permanent price impact eliminates round-trip arbitrage opportu-125
nities, as discussed in Appendix B.126
3.1 Geometric Brownian Motion127
In the GBM case, we assume the temporary price impact scales linearly with the asset price, i.e.128
Sexec(v, t) = f(v)S(t), (3.2)
where129
f(v) = [1 + κs sgn(v)] exp[κt sgn(v)|v|β ],
κs is the bid-ask spread parameter ,
κt is the temporary price impact factor,
β is the price impact exponent . (3.3)
Note that we assume κs < 1, so that Sexec(v, t) ≥ 0, regardless of the magnitude of v.130
4
3.2 Arithmetic Brownian Motion131
In the GBM case, we assume the temporary price impact is asset-price-independent, i.e.132
Sexec(v, t) = S(t) + S(0)h(v), (3.4)
where133
h(v) = κs sgn(v) + κtv , (3.5)
to be in accordance with Almgren and Chriss (2001). Note that Sexec(v, t) may be negative for v → −∞,134
i.e. (3.5) is only valid for small trading rates.135
Temporary impact (3.4), (3.5) is related to (3.2), (3.3) as follows. Assuming β = 1, κt|v| 1 and136
κtκs 1, temporary impact of the form (3.3) is approximately137
f(v) ≈ 1 + κs sgn(v) + κtv (3.6)
by Taylor’s expansion. Therefore, the price impact (in dollar terms) is138
Sexec(v, t)− S(t) = f(v)S(t)− S(t) = S(t)[κs sgn(v) + κtv] ≈ S(0)[κs sgn(v) + κtv]
for S(t) ≈ S(0).139
4 Definition of liquidation value140
Given the state (S(T−), A(T−)) at the instant t = T− before the end of the trading horizon, we have one141
final liquidation (if necessary) so that the number of shares owned at t = T is A(T ) = 0. The liquidation142
value B(T ) after this final trade is defined to be1143
B(T ) = B(T−) + limv→−∞
A(T−)Sexec(v, T−)
=
∫ T−
0
er(T−t′)(− vSexec(v, t′)
)dt′ + lim
v→−∞A(T−)Sexec(v, T
−). (4.1)
In the GBM case, applying (3.2) and (3.3) to (4.1) gives144
B(T ) = B(T−). (4.2)
In the ABM case, applying (3.4) and (3.5) to (4.1) gives145
B(T ) = B(T−) ; A(T−) = 0
= −∞ ; otherwise (4.3)
Definition (4.1) in effect penalizes the strategy if A(T ) 6= 0, so that the optimal algorithm forces the146
liquidation profile towards A(T ) = 0. In the GBM case (4.2), the penalty is such that the shares A(T−) are147
simply discarded. In the ABM case (4.3), a large penalty is imposed.2148
1Note that we adopt the convention that B(0) = 0; see (2.1).2In actual implementation, we would replace limv→−∞ by a finite vmin 0 in the PDE initial condition. Also, in the
case of liquidating a short position (buying), which is not considered in this paper, equation (4.1) would be defined as B(T ) =B(T−) + limv→∞ A(T−)Sexec(v, T−), and we would replace limv→∞ by a finite vmax 0 in implementation.
5
5 Mean Quadratic Variation formulation149
Quadratic variation has been used as an approximation of variance in the algorithmic trading literature150
(Almgren and Chriss, 2001; Lorenz and Almgren, 2006). This approximation, however, can be poor when151
the trading impact is relatively large, as explained in Appendix C and illustrated in Tse et al. (2011). Instead152
of using quadratic variation to approximate variance, it is conceptually simpler to regard quadratic variation153
as an alternative risk measure.154
5.1 Quadratic variation as a risk measure155
Formally, the quadratic variation risk measure is defined as156
E
[∫ T
t
(A(t′)dS(t′)
)2]. (5.1)
Informally, the risk measure definition (5.1) can be interpreted as the quadratic variation of the portfolio157
value process as follows: by expanding the square of dP (t′) = dB(t′)+d(A(t′)S(t′)) and ignoring higher-order158
terms, we have159 ∫ T
t
(A(t′)dS(t′)
)2=
∫ T
t
(dP (t′)
)2, (5.2)
when the trading velocity process v(t) is bounded.160
From the interpretation (5.2), minimizing quadratic variation clearly corresponds to minimizing volatility161
in the portfolio value process. The definition (5.1) shows that quadratic variation takes into account the162
trading trajectory A(t′) over the whole trading horizon. This is in contrast with using variance (V ar[B(T )])163
as a risk measure, which is independent of the trading trajectory A(t′) given the end result B(T ). We note164
that the idea of using quadratic variation as a risk measure was first suggested in Brugiere (1996).165
5.2 Objective functional and value function166
Now we specify the mean quadratic variation formulation as follows. For a fixed initial point (s, α, t) =167
(S(t), A(t), t) where t < T with B(t) = 0, trading strategy v(·), and risk aversion parameter λ, we define the168
objective functional169
J(s, α, t, v(·);λ) = Es,α,tv(·)
[B(T )
]− λEs,α,tv(·)
[∫ T
t
(A(t′)dS(t′)
)2](5.3)
where170
B(T ) =
∫ T−
t
er(T−t′)(− vSexec(v, t′)
)dt′ + lim
v→−∞A(T−)Sexec(v, T
−) (5.4)
and Es,α,tv(·) [·] is the conditional expectation at the initial point (s, α, t) using the control v(·).171
The value function V is defined as172
V (s, α, t;λ) = supv(·)
J(s, α, t, v(·);λ). (5.5)
For a given initial state (s, α, t), we will henceforth use the notation v∗s,α,t,λ(·) to denote the optimal policy173
that maximizes the corresponding functional, i.e. J(s, α, t, v(·);λ).174
6
5.3 Time Consistency of the optimal strategies175
Let (s1, α1, t1) be some state at time t1 and v∗s1,α1,t1,λ(·) be the corresponding optimal strategy. Let (s2, α2, t2)176
be some other state at time t2 > t1 and v∗s2,α2,t2,λ(·) be the corresponding optimal strategy.3177
Since the optimal controls satisfy the Bellman’s principle of optimality as shown in Appendix A, it follows178
that the optimal controls of (5.5) are time consistent in the sense that for the same state (s′, α′, t′) at a later179
time t′ > t2,180
v∗s1,α1,t1,λ(s′, α′, t′) = v∗s2,α2,t2,λ(s′, α′, t′) ; t′ ≥ t2 . (5.6)
In view of equation (5.6), we can drop the subscript and just write v∗(·).181
6 HJB Equation Formulation: GBM case182
For t < T , let V = V (s, α, τ = T − t;λ) = V (s, α, t;λ). For notational simplicity, we drop the parameter183
λ from V henceforth, i.e. we simply write V = V (s, α, τ). Unless otherwise stated, we also restrict the184
admissible controls v(·) to be non-positive (i.e. only selling is permitted).185
6.1 Optimal Control186
The optimal control v∗(·) can be obtained by solving the following HJB PDE derived in Appendix A:187
Vτ = µsVs +σ2s2
2Vss − λσ2α2s2 + sup
v≤0
[erτ (−vf(v))s+ g(v)sVs + vVα
]. (6.1)
Note that V has so far been defined for τ > 0 only. Section 4 suggests that the initial condition for188
V (s, α, τ = 0) should be189
V (s, α, τ = 0) = limv→−∞
αsf(v) (6.2)
for the GBM model with temporary price impact (3.2).190
6.2 Expected Value191
In order to construct the efficient frontier, i.e. a plot of expected gain versus risk, we will need to compute192
the expected gain. Let W (s, α, t) be the expected gain from the strategy v∗(·) found by solving equation193
(6.1), i.e.194
W (s, α, t) = Es,α,tv∗(·)
[B(T )
]. (6.3)
Let W (s, α, τ = T − t) = W (s, α, t), and following the same steps as used to derive equation (6.1), (or simply195
setting λ = 0), we obtain196
Wτ = µsWs +σ2s2
2Wss + erτ
(− v∗f(v∗)
)s+ g(v∗)sWs + v∗Wα . (6.4)
The initial condition W (s, α, τ = 0) is determined using the same arguments as used to derive (6.2)197
W (s, α, τ = 0) = limv→−∞
αsf(v). (6.5)
3Note that while the initial point is changed from (s1, α1, t1) to (s2, α2, t2), the risk aversion level λ is kept constant.
7
6.3 Construction of the Efficient Frontier198
For a given value of λ, we solve the nonlinear PDE (6.1), which gives us the optimal control v∗(·). With this199
optimal control v∗(·), we then solve the linear PDE (6.4). Let200
V0 = V (s = S(0), α = A(0), τ = T ) = Es,α,t=0v∗(·)
[B(T )− λ
∫ T
0
(A(t′)dS(t′)
)2],
W0 = W (s = S(0), α = A(0), τ = T ) = Es,α,t=0v∗(·)
[B(T )
]. (6.6)
In order to produce a plot of reward (expected) versus risk, we define risk so that it has the same dimensions201
as the expected gain, i.e.202
Risk =
(Es,α,t=0v∗(·)
[∫ T
0
(A(t′)dS(t′)
)2])1/2
=
√W0 − V0
λ. (6.7)
from equations (6.6). Equations (6.6) and (6.7) give us a single point on the efficient frontier. Repeating the203
above computation for different values of λ allows us to trace out the entire efficient frontier.204
6.4 Localization and Boundary Conditions205
6.4.1 Optimal Control Equation (6.1)206
The original problem (6.1) is posed on the domain207
(s, α, τ) ∈ [0,∞]× [0, αinit]× [0, T ] (6.8)
and we allow v to take any non-positive value. For computational purposes, we localize this domain to208
Ω = [0, smax]× [0, αinit]× [0, T ] (6.9)
and impose v ∈ [vmin, 0] for some finite negative value vmin.209
At α = 0, we do not permit selling which would cause α < 0, therefore v = 0 and hence210
Vτ = µsVS +σ2s2
2Vss ; α = 0, (6.10)
which does not require a boundary condition. Also, no boundary condition is required at α = αinit since211
v ≤ 0.212
At s = 0, no boundary condition is needed and we simply solve213
Vτ = supv∈[vmin,0]
[vVα
]; s = 0 (6.11)
At s = smax, we make the assumption that V ' C(α, τ)s2, which can be justified by noting that the term214
λσ2α2s2 acts as a source term in equation (6.1). We also assume that the effect of any permanent price215
impact at s = smax can be ignored i.e. g(v) = 0 at s = smax. This gives216
Vτ = (2µ+ σ2)V − λσ2α2s2 + supv∈[vmin,0]
[erτ (−vf(v))s+ vVα
]; s = smax . (6.12)
Equation (6.12) is clearly an approximation. We will carry out numerical tests with varying smax to show217
that the error in this approximation can be made small in regions of interest.218
The initial condition at τ = 0 is given by equation (6.2).219
8
6.4.2 Expected Value Equation (6.4)220
Similar to the situation for V , no boundary condition is needed for W at α = 0, α = αinit or s = 0. At221
s = smax, we assume W ' D(α, τ)s (based on the initial condition (6.5)) and g(v) = 0. Consequently,222
Wτ = µW + erτ (−v∗f(v∗))s+ v∗Wα ; s = smax. (6.13)
Again, equation (6.13) is clearly an approximation. We will verify that the effect of this is small for sufficiently223
large smax.224
The initial condition at τ = 0 is given by equation (6.5).225
7 HJB Equation Formulation: ABM case226
The derivation is similar to that in the GBM case. In this section, we show that by assuming asset-price-227
independent temporary price impact (3.4) and zero interest rate, the optimal strategy has no s-dependence.228
Under these assumptions, a derivation similar to that in Appendix A gives the HJB PDE229
Vτ = µS(0)Vs +σ2S(0)2
2Vss − λσ2α2S(0)2 + sup
v≤0
[v(Vα − s)− vh(v)S(0) + g(v)S(0)Vs
]. (7.1)
Note that the explicit s dependence in equation (7.1) appears only in the term v(Vα − s). Let230
U(s, α, τ) = V (α, s, τ)− αs. (7.2)
Substituting equation (7.2) into equation (7.1) gives231
Uτ = µS(0)(Us + α) +σ2S(0)2
2Uss − λσ2α2S(0)2 + sup
v≤0
[vUα − vh(v)S(0) + g(v)S(0)
(Us + α
)].
(7.3)
From equations (3.5) and (4.1), the initial condition for V is232
V (s, α, τ = 0) = limv→−∞
α(s+ S(0)h(v)). (7.4)
Therefore, from (7.2) we obtain233
U(s, α, τ = 0) = limv→−∞
α(s+ S(0)h(v))− αs = limv→−∞
αS(0)h(v) . (7.5)
Now, note that equation (7.3) has no explicit s dependence, and that the initial condition (7.5) has no s234
dependence. It therefore follows that equation (7.3) with initial condition (7.5) can be satisfied by a function235
U(α, τ) = U(s, α, τ) (7.6)
where U(α, τ) satisfies236
Uτ = µS(0)α− λσ2α2S(0)2 + supv≤0
[vUα − vh(v)S(0) + g(v)S(0)α
], (7.7)
with237
U(α, τ = 0) = limv→−∞
αS(0)h(v) . (7.8)
Proposition 1 Assuming Arithmetic Brownian Motion (2.6), asset-price-independent temporary price im-238
pact (3.4), zero interest rate, and initial condition (7.4), the optimal control for equation (7.1) is static, even239
when optimization is over the class of dynamic strategies.240
Proof . The optimal control for equation (7.1) is same as the optimal control for equation (7.7), which is241
independent of s, i.e. v∗(·) : (α, τ) 7→ v∗ , hence the optimal control for problem (7.1) is also independent of242
s, i.e. static. 243
9
7.1 Special case analytical solution of (7.7)244
In general, the PDE (7.7) has no known analytical solution. This section gives a special case analytical solu-245
tion under the additional assumptions of zero drift, unconstrained control and linear price impact functions.246
More formally, we make the following set of common assumptions 4247
Assumption 7.1
dS(t) = g(v)S(0)dt+ σS(0)dW(t),
r = 0,
h(v) = κs sgn(v) + κtv,
g(v) = κpv,
v ∈ (−∞,∞) (7.9)
which gives the following result.248
Proposition 2 Under Assumptions 7.1, the optimal control for (7.7) is identical with the (continuous equiv-249
alent of the) static strategy in Almgren and Chriss (2001); Almgren (2009), i.e.250
v∗(α, τ) = −αK cosh (Kτ)
sinh (Kτ)(7.10)
where K =√λσ2S(0)/κt.251
The value function U(α, τ) is252
U = E + λF , (7.11)
where253
E =S(0)α
(2κsf1(τ)2 + ακpf1(τ)2 + αλσ2S(0) τ + ακtKf1(τ)f2(τ)
)−2f1(τ)2
,
F =σ2S(0)
2α2(−f3(τ)2f1(τ)− f1(τ) + 2 τ Kf3(τ)
)4Kf1(τ)2f3(τ)
,
f1(τ) = sinh(Kτ) ,
f2(τ) = cosh(Kτ) ,
f3(τ) = exp(Kτ) .
(7.12)
Note that if κs = 0, then both E and F , and hence U are proportional to α2.254
Proof. Under Assumptions 7.1, the PDE (7.7) has the form255
Uτ = −λσ2α2S(0)2 + supv∈(−∞,∞)
[vUα −
(κsv sgn(v) + κtv
2)S(0) + κpvS(0)α
]. (7.13)
Using an initial condition that is consistent with (7.8) for U(α, τ) gives256
U(α, 0) = 0 ; α = 0
= −∞ ; otherwise , (7.14)
by using the definitions (7.2) and (7.6). It can be verified by straightforward calculations that the value257
function (7.11) and the control (7.10) solves the HJB PDE (7.13), (7.14). 258
In general, we would like to restrict v from taking all real values. For example, in the case of selling, a259
natural constraint is v ≤ 0 (the default in this paper). This constraint may take effect if µ 6= 0, in which260
case the analytical solution will no longer be valid.261
4Note that the assumption of unconstrained control may not be desirable as it allows buying shares during stock liquidation.
10
8 Approximations to pre-commitment mean variance262
As we mentioned in the introduction, the method typically used to evaluate the performance of an algorithmic263
execution strategy aligns well with the pre-commitment mean variance formulation. Here we give a brief but264
self-contained description of this formulation. We also discuss various approximations to this formulation.265
8.1 Pre-commitment mean variance266
For notational simplicity, we define x = (s, b, α) for a space state. Note that the state space is expanded to267
include b in this formulation, in contrast to the mean quadratic variation case. For a fixed initial point (x, t)268
where t < T , we define the functional269
JMV (x, t, v(·);λ) = Ex,tv(·)
[B(T )
]− λV arx,tv(·)
[B(T )
], (8.1)
where V arx,tv(·)[·] is the variance at the initial point (x, t) using the control v(·). Let (x0, 0) = (sinit, 0, αinit, 0)270
be the initial state. The corresponding optimal strategy v∗x0,0,λ(·) is termed the pre-commitment mean271
variance optimal strategy (Basak and Chabakauri, 2010). We note that the optimal strategy v∗x0,0,λ(·) is272
dynamic, in both the GBM case (Forsyth, 2011) and the ABM case (Lorenz and Almgren, 2006; Lorenz,273
2008).274
The pre-commitment strategy is optimal in the following sense: suppose we carry out many thousands275
of trades. We then examine the post-trade data, and determine the realized mean return and the standard276
deviation. Assuming that the modeled dynamics very closely match the dynamics in the real world, the277
optimal pre-commitment strategy would result in the largest realized mean return, for a given standard278
deviation, compared to any other possible strategy.279
Although the pre-commitment mean variance formulation is consistent with evaluation of performance280
of algorithmic trading trading strategies in practice, these optimal strategies are time-inconsistent (Basak281
and Chabakauri, 2010; Wang and Forsyth, 2010; Forsyth, 2011; Wang and Forsyth, 2011a), a property that282
is considered unnatural by some authors (Basak and Chabakauri, 2010).283
8.1.1 Time-inconsistency of optimal strategies284
Let (x1, t1) be some state at time t1 and v∗x1,t1,λ(·) be the corresponding optimal policy. Let (x2, t2) be some285
other state at time t2 > t1 and v∗x2,t2,λ(·) be the corresponding optimal policy. We have time-inconsistency286
in the sense that287
v∗x1,t1,λ(x′, t′) 6= v∗x2,t2,λ(x′, t′) ; t′ ≥ t2 . (8.2)
As discussed in Basak and Chabakauri (2010), there is no direct dynamic programming principle for288
determining v∗x0,0,λ(·) due to time-inconsistency. Forsyth (2011) uses a Lagrange multiplier method to solve289
for v∗x0,0,λ(·).290
We now discuss various approximations to the pre-commitment mean variance problem that lead to291
time-consistent optimal strategies.292
8.2 Approximation 1: Restrict to static strategies293
The Almgren and Chriss (2001) approximation essentially restricts the admissible strategies to static strate-294
gies in optimizing (8.1). As discussed in the previous section, this is suboptimal, in both the GBM and295
the ABM case. It is interesting that, for this approximation problem of maximizing the mean variance296
functional (8.1) over static strategies (assuming ABM), the optimal strategies are time-consistent. This297
time-consistency can be verified using the formula (7.10).298
11
8.3 Approximation 2: Use quadratic variation as risk measure299
Another approach is to approximate variance by quadratic variation, the accuracy of which is discussed in300
Appendix C. As discussed in Section 5.1, quadratic variation can be justified as a reasonable alternative301
risk measure. The current paper studies this approximation. Our more recent paper (Tse et al., 2011)302
compares optimal strategies in this approximation to the truly optimal strategies for the pre-commitment303
mean variance formulation. The time-consistency of this formulation is discussed in Section 5.3.304
8.4 Approximation 3: Restrict to time-consistent strategies305
If the only criticism of the pre-commitment mean variance formulation is that its optimal strategies are306
time-inconsistent, the optimization can be restricted to optimizing over time-consistent strategies. This is307
in some sense similar to Approximation 1, which optimizes over static strategies. However, the restriction308
to time-consistent strategies is more subtle, as we explain below.309
The restriction to static strategies is easy to understand since we can look at a single strategy v(·) and310
determine whether it is static or not. This is not the case for the restriction to time-consistent strategies.311
It is important to note that time-consistency concerns the relation between a continuum of strategies, as312
explained in Section 5.3 and 8.1.1, and cannot be inferred from examination of a single strategy v(·).313
Although it is more difficult to enforce time-consistency as a restriction (in terms of defining the cor-314
responding class of admissible strategies), it turns out that this is not necessary to solve for the optimal315
time-consistent solutions. Essentially, since the time-consistent strategies follow the Bellman’s principle of316
optimality by definition, the dynamic programming principle can be used to solve for the optimal time-317
consistent solutions. We refer readers to Basak and Chabakauri (2010) for details.318
From a computational perspective, the optimal time-consistent strategies are in fact more difficult to319
determine (Wang and Forsyth, 2011b) compared to the pre-commitment optimal strategies. We also note320
that pre-commitment and time consistent strategies are the same as T → 0 (Basak and Chabakauri, 2010).321
In some special cases, the optimal time consistent strategies are identical to the optimal strategies in the322
mean-quadratic variation formulation (Bjork et al., 2009).323
8.5 Connection between Approximation 1 and 2324
Approximation 1 restricts attention to static strategies. Approximation 2 approximates variance by quadratic325
variation. These two approximations have the connection that variance is the same as (the expected value326
of) quadratic variation for static strategies, i.e.327
V arx,t=0v(·)
[B(T )
]= Ex,t=0
v(·)
[∫ T
0
(A(t′)dS(t′))2]
(8.3)
under some additional mild assumptions detailed in Appendix C.328
Nevertheless, we emphasize that this equality, by itself, does not imply that the static solution in Almgren329
and Chriss (2001) is optimal for the quadratic variation risk measure. The subtle point is that Almgren and330
Chriss (2001) does not study how dynamic strategies (i.e. use information regarding how the stock price331
evolves after the start of trading) perform in terms of the quadratic variation risk measure. In this paper,332
we provide a proof of this optimality. Recall that Section 7.1 shows that the strategy (7.10) solves (5.5),333
while Proposition 1 shows that no dynamic strategy is better than this static strategy.334
In general, the equality (8.3) does not hold, and quadratic variation is only an approximation to variance.335
Although the accuracy of this approximation is good when trading impact is small (compared to volatility),336
this approximation can be poor when trading impact is larger but still realistic. Our proof in Appendix C337
shows precisely what is ignored in this approximation. Examples in which the approximation is poor can be338
found in Tse et al. (2011).339
Remark 8.1 (Static as a restriction or as a property) It should be clear at this point that it is impor-340
tant to distinguish between whether the use of the concepts static or dynamic refer to the class of admissible341
12
strategies, or to a property of the optimal control. In particular, even if the class of admissible controls is342
dynamic, the optimal control may turn out to be static. Proposition 1 is an example. Another example can343
be found in Schied et al. (2010).344
9 Numerical Method: GBM Case (6.1)345
We give a brief outline of the numerical method used to solve equation (6.1). We will use a semi-Lagrangian346
method, similar to the approach used in Chen and Forsyth (2007).347
Along the trajectory s = s(τ), α = α(τ) defined by348
ds
dτ= −g(v)s
dα
dτ= −v , (9.1)
equation (6.1) can be written as349
supv≤0
[DV
Dτ(v)− LV − erτ (−vf(v))s− λα2s2σ2
]= 0 , (9.2)
where the operator LV is given by350
LV = µsVs +σ2s2
2Vss , (9.3)
and where the Lagrangian derivative DVDτ (v) is given by351
DV
Dτ(v) = Vτ − Vsg(v)s− Vαv . (9.4)
The Lagrangian derivative is the rate of change of V along the trajectory (9.1).352
Define a set of nodes [s0, s1, ..., simax], [α0, α1, ..., αkmax
], discrete times τn = n∆τ , and localize the control353
candidates to values in a finite interval [vmin, vmax]. Let V (si, αj , τn) denote the exact solution to equation354
(6.1) at point (si, αj , τn). Let V ni,j denote the discrete approximation to the exact solution V (si, αj , τ
n).355
We use standard finite difference methods (d’Halluin et al., 2005) to discretize the operator LV as given356
in (9.3). Let (LhV )ni,j denote the discrete value of the differential operator (9.3) at node (si, αj , τn). The357
operator (9.3) can be discretized using central, forward, or backward differencing in the s direction to give358
(LhV )ni,j = aiVni−1,j + biV
ni+1,j − (ai + bi)V
ni,j , (9.5)
where ai and bi are determined using the algorithm in d’Halluin et al. (2005).359
Let vni,j denote the approximate value of the control variable v at mesh node (si, αj , τn). Then we360
approximate DVDτ (v) at (si, αj , τ
n+1) by the following361 (DV
Dτ(v)
)n+1
i,j
' 1
∆τ(V n+1i,j − V n
i,j) (9.6)
where V ni,j
is an approximation of V (sni, αn
j, τn) obtained by linear interpolation of the discrete values V ni,j ,362
with (sni, αn
j) given by solving equations (9.1) backwards in time for fixed vn+1
i,j to give363
sni
= si + sig(vn+1i,j )∆τ +O(∆τ)2
αnj
= αj + vn+1i,j ∆τ . (9.7)
13
Our final discretization is then364
V n+1i,j = sup
vn+1i,j ∈[vmin,vmax]
[V ni,j
+ ∆τ erτn+1
(−vn+1i,j f(vn+1
i,j ))
]+ ∆τ(LhV )n+1
i,j −∆τλ(αj)2s2iσ
2 .
(9.8)
Let365
∆smax = maxisi+1 − si
∆αmax = maxjαj+1 − αj , (9.9)
and define a discretization parameter h such that366
h =∆αmax
C1=
∆smax
C2=
∆τ
C3(9.10)
where Ci are positive constants. Note that we must solve a local optimization problem at each node at each367
time step in equation (9.8). In fact, we need to determine the global maximum of the local optimization368
problem. If the set of controls [vmin, vmax] is discretized with spacing h, then a linear search of the control369
space will converge to the viscosity solution of the HJB equation (6.1) (Wang and Forsyth, 2008). An370
alternative (and less computationally expensive) method is to use a one dimensional optimization algorithm371
(Brent, 1973) to determine the local optimal control. The difficulty here is that one dimensional optimization372
methods are not guaranteed to converge to the global maximum. We will carry out tests using both methods373
in the following.374
10 Numerical Method ABM Case (7.7)375
Similar to the derivation in the last section, (7.7) can be written as376
supv≤0
[DU
Dτ(v) + vh(v)S(0)− g(v)S(0)α
]= µS(0)α− λα2S(0)2σ2 (10.1)
where the Lagrangian derivative DUDτ (v) = Uτ − Uαv.377
By integrating along the Lagrangian path and discretizing, we obtain378
Un+1j = sup
vn+1j ∈[vmin,vmax]
(Unj
+ ∆τ(g(vn+1
j )S(0)αj − vn+1j h(vn+1
j )S(0)))
+ ∆τ
(µS(0)αj − λα2
jS(0)2σ2
),
(10.2)
where we have used the notation Un+1j = U(αj , τ
n+1), vn+1j = vj(τ
n+1) and Unj≈ U(αn
j, τn). Here αn
jis379
defined as in equation (9.7).380
Either linear or quadratic interpolation can be used in approximating Unj
. Linear schemes have the381
advantage that they are monotone and numerical solutions are guaranteed to converge to the viscosity382
solution of the HJB equation, whereas quadratic schemes may not converge to the viscosity solution. In the383
special case where the analytical solution (7.12) is known, our quadratic interpolation scheme does converge384
to the exact solution of the HJB equation (10.1).385
11 Numerical Examples386
We solve both the GBM problem (6.1) and the ABM problem (7.1). Recall that the efficient frontier is387
constructed as described in Section 6.3, where we define388
Risk =
(Es,α,t=0v∗(·)
[∫ T
0
(A(t′)dS(t′)
)2])1/2
(11.1)
14
Parameter Valueσ .40T 1/12µ 0.0r 0.0S(0) 100A(0) 1.0Action Sellκp 0.0κt .002κs 0.0β 1.0vmin −105/Tvmax 0.0Smax 5000
Table 1: Parameters for Case 1: selling an illiquid stock over a long trading horizon.
Refinement Time steps S nodes α nodes v nodes0 100 67 41 301 200 133 81 592 400 265 161 1173 800 529 321 2334 1600 1057 641 465
Table 2: Grid and time step information for various levels of refinement for parametric case in Table 1.
to have the same units as expected gain.389
We will consider two cases. Case 1 considers an illiquid stock traded over a long time horizon (one month).390
We only consider the GBM model because the ABM model is unrealistic in this case. Case 2 considers a391
liquid stock traded over a short time horizon (one day). We consider both the GBM and the ABM model in392
this case and compare the results.393
11.1 Example 1: Illiquid Stock, Long Trading Horizon (GBM)394
The parameters for this case are shown in Table 1. The value of κt in Table 1 corresponds to a temporary395
price impact of about 240 bps for liquidating at a constant rate over the entire month. This would correspond396
to an illiquid stock. A relatively large volatility is also assumed, the value of σ in Table 1 corresponds to a397
standard deviation of about 1154 bps of S(T ).398
11.1.1 Convergence Tests399
We will first carry out some convergence tests, using the data in Table 1. The grid and time step information400
are given in Table 2401
As noted in Section 9, in general, we need to use a linear search to guarantee that the global maximum402
of the local optimization problem at each node in equation (9.8) is determined to O(h) for smooth functions.403
This guarantees convergence to the viscosity solution of equation (6.1).404
Tables 3 and 4 compare results using a linear search or a one dimensional optimization technique to405
solve the local optimization problem at each node. These tables clearly show that both methods converge406
to the same solution. We have verified that the one dimensional optimization method converges to the407
15
Refinement Value Function Expected Gain Risk Control0 91.8439869 95.80303404 4.449183728 -40.51 91.9609793 95.85387412 4.411856106 -41.252 92.0205568 95.88081381 4.39332277 -41.6253 92.0510158 95.88846915 4.38032725 -41.8125
Table 3: Convergence test for using linear search of discrete trade rates for parametric case shown in Table1. The grid and time step information for each refinement is given in Table 2. All values are reported ats = S(0) = 100, α = A(0) = 1, τ = T for λ = 0.2. Value function is V as defined in (6.1). Expected gain isW as defined in (6.6). Risk is
√(W − V )/λ as defined in (6.7). Control is v∗ as defined in (6.4), which is
determined using a linear search. Compare with Table 5 which shows the result of using a one dimensionaloptimization method to determine v∗.
Refinement Value Function Expected Gain Risk Control0 91.8466222 95.80519281 4.448915945 -41.09866951 91.9616700 95.85496493 4.412082788 -41.48291342 92.0207373 95.88384474 4.394944527 -41.66571673 92.0510986 95.90026189 4.387005382 -41.7545063
Table 4: Convergence test for using a one dimensional optimization method for parametric case shown inTable 1. Notations are as in Table 3 and results are computed in the same way except that the control isdetermined using a one dimensional optimization method. Note that the values in the two tables appear toconverge to the same limit.
global optimum in many tests, which we will not report. In the following, we will use the one dimensional408
optimization method, since it is much less expensive, computationally.409
Recall that we made several approximations in order to determine boundary conditions at S = Smax.410
Table 5 shows the effect of increasing Smax, and verifies that the effect of these boundary condition approx-411
imations is negligible in regions of interest.412
Another test of convergence is to consider the special case analytic solution for λ = 0. In this case where413
expected gain is maximized regardless of risk, the optimal selling strategy should sell at a constant rate to414
minimize temporary trading impact. This can be proved by noting the parametric choice µ = r = 0 in415
Table 1 and the form of temporary price impact (3.3). Since µ = r = 0, and the constant liquidation rate is416
v = −1/T = −1/12, the expected gain will be (using the parameters in Table 1)417
E[e−κt/T
1
T
∫ T
0
S(t) dt]' 97.6286 (11.2)
Table 6 shows the results for the expected gains for λ = 0.0001. The table shows that the numerical results418
appear to be converging to the analytic solution for λ = 0.419
Refinement Value Function Expected Gain Risk Control0 91.8466222 95.80519281 4.448915945 -41.09866951 91.9616700 95.85496493 4.412082788 -41.48291342 92.0207373 95.88384474 4.394944527 -41.6657167
Table 5: Test to confirm increasing Smax makes no difference for the parametric case as shown in Table1. Notations are as in Table 4 and results are computed in the same way, except that this table usesSmax = 20000 instead of Smax = 5000. Note that there is negligible difference between the two tables.
16
Refinement Expected Gain0 97.48419611 97.54478892 97.58079203 97.60170004 97.6136413
Table 6: Convergence to analytical solution for constant liquidation rate for parametric case shown inTable 1. The grid and time step information for each refinement is given in Table 2. Expected gains W ,as defined in (6.6), are reported at s = S(0) = 100, α = A(0) = 1, τ = T for λ = 0.0001 ≈ 0. Optimalcontrol is determined using a one dimensional optimization method. Expected gains appear to converge tothe analytical value 97.6286.
11.1.2 Efficient Frontier420
Figure 1 shows the efficient frontier traced out by the optimal trading strategies with different risk aversion421
level λ. Note that even the coarsest grid gives accurate results for expected gain values of interest.422
11.1.3 Optimal Trading Rates423
Figure 2 shows the optimal trading rate v∗(s, α, t;λ) as a function of asset price s. Note that the optimal424
strategy is slightly aggressive-in-the-money, i.e. the optimal strategy is to sell faster at larger s.425
11.1.4 Comparison with pre-commitment mean-variance solution426
The parametric case as shown in Table 1 is also studied in Forsyth (2011) which solves for the optimal427
strategies for the pre- commitment mean variance formulation. Here we briefly compare the two formulations428
from the perspective of numerical solution for the optimal controls. For more comprehensive comparison429
between the two formulations, we refer readers to (Tse et al., 2011).430
It is more difficult to numerically solve for the optimal strategies in the pre-commitment mean variance431
formulation. Essentially this is because the local objective as a function of trading rate is sometimes very432
flat, making it difficult to determine the maximizer numerically. This flatness of the objective function is433
related to the fact that the variance risk measure does not concern the trading trajectory A(t′) but only434
the end result B(T ) as we discussed in Section 5.1. In contrast, this flatness is not observed in the mean435
quadratic variation formulation.436
The numerical difficulties encountered in the pre-commitment mean variance formulation can be seen437
from two aspects. First, while both one dimensional optimization and linear search are able to find the438
maximizer in the mean quadratic variation formulation, as we discussed previously in this section, one439
dimensional optimization does not work well for the pre-commitment mean variance formulation and hence440
the more computationally expensive linear search method needs to be used. Second, in the pre-commitment441
mean variance formulation the optimal trading rate (as a function of asset price) is oscillatory, which reflects442
the near ill-posedness of this formulation (i.e. there are many strategies which give almost the same mean443
and variance). In contrast, the optimal trading rate is smooth in the mean quadratic variation formulation,444
as shown in Figure 2.445
11.2 Example 2: Liquid Stock, Short Trading Horizon446
The parameters for this case are shown in Table 7. The value of κt in Table 7 corresponds to a temporary447
price impact of about 5 bps for liquidating at a constant rate over the trading day. This would correspond to448
a liquid stock. The larger value of σ = 1.0 in Table 7 corresponds to a standard deviation of about 632 bps449
of S(T ). Therefore, this parametric case considers a situation where volatility is large compared to trading450
impact.451
17
Risk
Exp
ecte
dG
ain
0 1 2 3 4 5 6 70
10
20
30
40
50
60
70
80
90
100
Refinement = 0
Refinement = 1
Refinement = 2
Figure 1: The efficient frontier for parametric case shown in Table 1. The grid and time step informationfor each refinement is given in Table 2. Values of expected gain, as defined in (6.6), and risk, as defined in(6.7) are reported at s = S(0) = 100, α = A(0) = 1, τ = T for various risk aversion level λ. Smaller valuesof λ represent less risk-averse strategies which have larger risks and expected gains.
Asset Price
Tra
deR
ate
0 50 100 150 200-70
-60
-50
-40
-30
-20
-10
0
Refinement = 1
Refinement = 2
Figure 2: Optimal trading rate v∗(s, α, t;λ) as a function of s at t = 0, α = 1, and λ = 0.2. The riskaversion level λ = 0.2 corresponds to the point on the efficient frontier in Figure 1 with expected gain95.9 and risk 4.4. Compare the trading rates with the constant liquidation rate v = −12. This is for theparametric case shown in Table 1. The grid and time step information for each refinement is given in Table2.
18
Parameter Valueσ 1.0T 1/250κt 2× 10−6
Table 7: Parameters for Case 2: selling a liquid stock over a short trading horizon. Other parameters areas given in Table 1.
Refinement Time steps S nodes α nodes v nodes0 800 67 41 301 1600 133 81 592 3200 265 161 1173 6400 529 321 233
Table 8: Grid and time step information for various levels of refinement for parametric case in Table 7.
For both the GBM case and the ABM case, the grid and time step information are given in Table 8.452
Recall that in the GBM case, we solve equation (6.1); in the ABM case, we solve equation (7.7).453
11.2.1 Geometric Brownian Motion454
The efficient frontier for the GBM case is shown in Figure 3. Note that even the coarsest grid gives accurate455
results for expected gain values of interest, similar to the parametric case in illiquid stock, long trading456
horizon case in Table 1. Convergence results are shown in Table 9 for various values of λ. We note that457
convergence appears to be at a first order rate. The optimal trading rates as a function of asset price are458
shown in Figure 3 for various values of λ. It shows that the optimal strategies are more aggressive-in-the-459
money (slope of the curves are larger) for larger values of λ.460
11.2.2 Arithmetic Brownian Motion461
The ABM case also uses the parameters in Table 7 with the form of temporary pricing impact changed to462
(3.5). Convergence results are shown in Table 10. Note that the numerical values appear to converge to the463
analytical solution. Optimal trading rates are not plotted for the ABM since they are independent of the464
asset price s and can be obtained in Table 10.465
11.2.3 Using optimal strategies from ABM as approximate solutions for GBM dynamics466
Section 11.2.1 assumes the stock price process follows GBM and solves for the optimal strategies, which are467
dynamic. Section 11.2.2 assumes the stock price process follows ABM and solves for the optimal strategies,468
which turn out to be static. In this section, we compare the performance of the strategies in these two cases,469
assuming the stock price process follows GBM and the temporary price impact is of the form (3.2), as in470
Section 11.2.1. Note that by making these assumptions, the GBM strategies are truly optimal whereas the471
ABM strategies are not. The reason for conducting this comparison is that the static strategies in 11.2.2472
have analytical solutions which can be considered as easy-to-compute approximate solutions to the optimal473
dynamic strategies in 11.2.1.474
This comparison is shown in Figure 5 where we compare the efficient frontiers obtained by the truly475
optimal dynamic strategies and the approximate static strategies. In Figure 5, the frontier labeled with476
“Exact Control” is the same as in Figure 3, the frontier labeled with “Approximate Control” is generated477
using the static strategy approximation. Surprisingly, the frontier generated by the approximate solution478
is virtually identical with the truly optimal one. This indicates that there is essentially no error, as far as479
19
Risk
Exp
ecte
dG
ain
0 1 2 3 490
91
92
93
94
95
96
97
98
99
100
Refinement = 0
Refinement = 1
Figure 3: The efficient frontier for parametric case shown in Table 7 in the GBM case. The grid and timestep information for each refinement is given in Table 8. Values of expected gain and risk are reported ats = S(0) = 100, α = A(0) = 1, τ = T for various risk aversion level λ. Smaller values of λ represent lessrisk-averse strategies which have larger risks and expected gains.
Refinement Expected Gain Risk Controlλ = 100
1 92.942727 .284003 -69312.32 92.926803 .271555 -72483.63 92.925507 .265444 -74491.4
λ = 101 97.756612 .482818 -22205.42 97.602898 .476058 -22552.93 97.620839 .472658 -22718.4
λ = 11 99.287578 .847254 -7058.172 99.290223 .843065 -7090.363 99.291576 .840960 -7106.59
λ = .21 99.68109 1.26727 -3140.722 99.68241 1.26168 -3156.803 99.68308 1.25887 -3164.78
Table 9: Convergence test for using one dimensional optimization for parametric case shown in Table 7 inthe GBM case. The grid and time step information for each refinement is given in Table 8. Notations areas in Table 3. All values are reported at s = S(0) = 100, α = A(0) = 1, τ = T for various values of λ. Notethat convergence appears to be at first order rate. Compare the trading rates with the constant liquidationrate v = −250.
20
Asset Price
Tra
deR
ate
0 50 100 150 200
-30000
-20000
-10000
0 λ = 0.2
λ = 1.0
λ = 10
Figure 4: Optimal trading rate v∗(s, α, t;λ) as a function of s at t = 0 and α = 1 for various values of λ.This is for the parametric case shown in Table 7 in the GBM case. The grid and time step information foreach refinement is given in Table 8. Compare the trading rates with the constant liquidation rate v = −250.
Refinement Expected Gain Risk Controlλ = 100
1 92.336427 .304879 -70000.02 92.612649 .284365 -70341.53 92.770613 .274943 -70588.9
analytic 92.928932 .265915 -70710.7λ = 10
1 97.694545 .493519 -22312.02 97.728951 .483026 -22352.63 97.746400 .477908 -22359.0
analytic 97.763932 .472871 -22360.7λ = 1
1 99.281307 .853065 -7070.302 99.287096 .846951 -7070.883 99.289994 .843916 -7071.02
analytic 99.292893 .840896 -7071.07λ = .2
1 99.678547 1.267777 -3174.842 99.681165 1.262605 -3168.523 99.682470 1.260019 -3165.41
analytic 99.683772 1.257433 -3162.28
Table 10: Convergence test for parametric case shown in Table 7 in the ABM case. The grid and time stepinformation for each refinement is given in Table 8. Notations are as in Table 3. All values are reportedat s = S(0) = 100, α = A(0) = 1, τ = T for various values of λ. Note that the numerical results appear toconverge to the analytical solution. Compare the trading rates with the constant liquidation rate v = −250.
21
Risk
Exp
ecte
dG
ain
0 1 2 3 490
91
92
93
94
95
96
97
98
99
100
Approximate Control
Exact Control
Figure 5: Comparison of the efficient frontier computed by solving equation (6.1) (exact control) comparedto constructing the efficient frontier using the control from the ABM approximation (7.7) (approximatecontrol). The asset price process follows GBM and the temporary price impact takes the form (3.2) incomputing both efficient frontiers.
the efficient frontier is concerned, even though the static strategies trade at a different rate than the truly480
optimal dynamic strategies trade.481
The accuracy of the static ABM strategies in approximating the dynamic GBM strategies can be explained482
as follows. At higher level of expected gains in Figure 5, both strategies must sell near the constant liquidation483
rate (throughout the entire trading horizon) to reduce trading impact, and therefore are similar. At lower484
level of expected gains, both strategies sell very quickly near t = 0, and actually finish most of the liquidation485
before the share price has time to move significently, i.e. most liquidation happens at S(t) ≈ S(0) = 100.486
Note that for the same level of expected gain, the two strategies trade at similar rates5 for s near 100, which487
implies that they trade at similar rates near t = 0. Consequently, the ABM approximation is also accurate488
for lower level of expected gain.489
The observation above suggests that by increasing only volatility (to values much larger than σ = 1490
considered here), the ABM approximation will still give efficient frontiers that are close to optimal for491
all risk levels. However, if volatility and temporary trading impacts are both increased substantially (to492
unrealistic levels), then we will see the sub-optimality of the ABM approximation at lower level of expected493
gains, though it is still accurate at higher level of expected gains. These conjectures are indeed confirmed494
by our numerical results, which are not reported here due to the unrealistic parameter values.495
The above discussion may give the wrong impression that it suffices to have the correct trading rate at496
s = 100 and being aggressive-in-the-money has no advantage. We emphasize, however, that this an over-497
simplification that is not true in general. When risk is measured by variance, instead of quadratic variation,498
being aggressive in the money can reduce risk substantially (Tse et al., 2011). In other words, the above499
simplification serendipitously happens to work when risk is measured by quadratic variation. This illustrates500
that quadratic variation and variance are different risk measures that lead to different optimal strategies, as501
we point out in Section 5.1 and elaborated in (Tse et al., 2011).502
5However, the difference increases as s moves away from 100. See Figure 4.
22
12 Conclusion503
We have proposed a mean–quadratic-variation objective function for determining the optimal trade execution504
strategy. Quadratic variation as a risk measure takes account of the entire trading trajectory. This is in505
contrast with using variance as a risk measure which only considers the terminal portfolio value distribution.506
The static strategy in Almgren and Chriss (2001), which is originally derived as an approximate solution507
to the pre-commitment mean variance problem, turns out to be the truly optimal solution in the mean508
quadratic variation formulation (assuming ABM).509
We have developed numerical schemes for solution of the mean–quadratic-variation optimal control prob-510
lem, assuming either GBM or ABM. Any type of constraint can be imposed on the trading strategy. For511
example, the natural constraint when selling is that no intermediate buying is allowed.512
In the GBM case, the optimal strategy depends smoothly on the underlying asset price. Numerical513
difficulties seen in the pre-commitment mean-variance formulation (Forsyth, 2011) are not seen in the mean514
quadratic variation formulation. In the ABM case, the optimal strategies are static, and thus different515
from those in the GBM case. Surprisingly, it turns out that the static strategies can be used as excellent516
approximations for the GBM case, even when volatility is large. We note that this accuracy of the static517
approximation does not hold in the pre-commitment mean variance formulation.518
Finally, we emphasize that in general, mean quadratic variation and mean variance are not the same519
objective functions, and that the optimal strategies in each case can be significantly different. However,520
there are arguments to made for choosing each of these objective functions.521
A Optimal Control522
In this appendix, we give the steps used to derive equation (6.1). Using equation (2.5), then the risk term523
becomes524 ∫ T
t
(A(t′) dS(t′))2
=
∫ T
t
σ2A(t′)2S(t′)2 dt′ , (A.1)
so that by equations (5.5) and (5.3) we have (using Es,α,tv(·)[Es+∆s,α+∆α,t+∆tv(·) [·]
]= Es,α,tv(·) [·])525
V (s, α, t;λ) = supv(·)
Es,α,tv(·)
[B(T )− λ
∫ T
t
(A(t′) dS(t′))2]
= supv(·)
Es,α,tv(·)
[∫ T
t
[er(T−t
′)(− vSexec(v, t′)
)− λσ2A(t′)2S(t′)2
]dt′ + lim
v→−∞A(T−)Sexec(v, T
−)
]
= supv(·)
Es,α,tv(·)
[∫ t+∆t
t
[er(T−t
′)(− vSexec(v, t′)
)− λσ2A(t′)2S(t′)2
]dt′
+ Es+∆s,α+∆α,t+∆tv(·)
∫ T
t+∆t
[er(T−t
′)(− vSexec(v, t′)
)− λσ2A(t′)2S(t′)2
]dt′
+ Es+∆s,α+∆α,t+∆tv(·)
[lim
v→−∞A(T−)Sexec(v, T
−)
]](A.2)
Noting that for any control v(·) : (S(t′), A(t′), t′) 7→ v, t′ ≥ t+ ∆t,526
Es+∆s,α+∆α,t+∆tv(·)
[∫ T
t+∆t
[er(T−t
′)(− vSexec(v, t′)
)− λσ2A(t′)2S(t′)2
]dt′ + lim
v→−∞A(T−)Sexec(v, T
−)
]= J(s+ ∆s, α+ ∆α, t+ ∆t, v(·);λ)
≤ supv(·)
J(s+ ∆s, α+ ∆α, t+ ∆t, v(·);λ) = V (s+ ∆s, α+ ∆α, t+ ∆t;λ). (A.3)
23
with equality in the case of the optimal control v∗(·).527
From equations (A.2-A.3) and the form of the price impact (3.2),528
V (s, α, t;λ) = supv(·)
Es,α,tv(·)
[er(T−t)(−vf(v)s)∆t− λσ2α2s2∆t
+V (s+ ∆s, α+ ∆α, t+ ∆t;λ)
]+O((∆t)2). (A.4)
Defining529
∆V = V (s+ ∆s, α+ ∆α, t+ ∆t;λ)− V (s, α, t;λ), (A.5)
and rearranging equation (A.4) gives530
0 = supv(·)
Es,α,tv(·)
[er(T−t)(−vf(v)s)∆t− λσ2α2s2∆t+ ∆V
]+O((∆t)2) . (A.6)
From equations (2.3) and (2.5), using Ito’s Lemma we obtain531
Es,α,tv(·) [∆V ] = ∆t
[Vt + (µ+ g(v)sVs +
σ2s2
2Vss + vVα
]+O((∆t)3/2) . (A.7)
Let V = V (s, α, τ = T − t;λ) = V (s, α, t;λ). Substituting equation (A.7) into equation (A.6), dividing532
by ∆t, and letting ∆t→ 0, then we obtain the HJB PDE for τ > 0533
Vτ = µsVs +σ2s2
2Vss − λσ2α2s2 + sup
v
[erτ (−vf(v))s+ g(v)sVs + vVα
]. (A.8)
B Form of Permanent Price Impact (3.1)534
Since temporary impact always leads to trading losses, there is no restriction on the functional form of535
the temporary impact (Huberman and Stanzl, 2004; Almgren et al., 2004). In contrast, the form of the536
permanent price impact must be restricted to ensure no-arbitrage, as noted by Huberman and Stanzl (2004).537
In this Appendix, we show that a permanent price impact function of the form (3.1) is consistent with the538
no-arbitrage condition of Huberman and Stanzl (2004), which basically states that the expected gain from539
a round trip trading strategy should be non-positive.540
Note that while previous work considered only Arithmetic Brownian Motion, we handle both cases here.541
Since the proofs are very similar, we only give full details for the GBM case here.542
In the following, we assume that there is no temporary price impact, since this is irrelevant in terms of no-543
arbitrage. Further, we assume that the deterministic drift term µ = 0 in equation (2.5), and g(v(t)) = κpv(t),544
with κp = constant.545
Consequently, we consider a process of the form546
dS(t) = κpv(t)S(t)dt+ σS(t)dW(t) . (B.1)
The solution of this SDE is547
S(T ) = S(0) exp
[κp
∫ T
0
v(t) dt
]exp
[σW(T )− σ2T/2
]. (B.2)
Noting that548 ∫ T
0
v(t) dt =
∫ T
0
dA(t)
dtdt =
∫ T
0
dA(t) = A(T )−A(0) (B.3)
24
then equation (B.2) becomes549
S(T ) = S(0) exp
[κp(A(T )−A(0))
]exp
[σW(T )− σ2T/2
], (B.4)
and consequently550
E[S(T )] = S(0) exp
[κp(A(T )−A(0))
]. (B.5)
For a round trip trade, A(T ) = A(0), hence551
E[S(T )] = S(0) . (B.6)
Let R(t) be the revenue from a trading strategy v(t), so that552
dR(t) = −v(t)S(t) dt. (B.7)
Rearranging equation (B.1), we obtain553
v(t)S(t) dt =dS(t)
κp− σS(t)
κpdW(t) . (B.8)
Substituting equation (B.8) into (B.7) gives554
R(T ) = −∫ T
0
[dS(t)
κp− σS(t)
κpdW(t)
]= −S(T )− S(0)
κp+
σ
κp
∫ T
0
S(t) dW(t) . (B.9)
Noting that555
E[∫ T
0
S(t) dW(t)]
= 0 , (B.10)
then, for a round trip trade (from equation (B.6))556
E[R(T )] = −E[S(T )]− S(0)
κp= 0 . (B.11)
Consequently, the expected revenue for any round trip trade for a permanent price impact of the form (3.1)557
is zero, hence this precludes arbitrage. Note that equation (B.6) also holds in the ABM case, and the rest558
of the proof is similar.559
C Derivation of Equation (8.3)560
In this Appendix, we reconstruct the arguments used to derive equation (8.3). The reader should note the561
following assumptions:562
AS1 The underlying process S(t) has no drift and563
dS(t′) = σ(S(t′), t′) dW(t′) . (C.1)
AS2 The control v(·) is of the form v(·) : (S(t), A(t), t) 7→ v.564
AS3 r = 0, A(T−) = A(T ) = 0, and the temporary impact is of the form (3.4)565
25
From equations (2.4), (3.4) we have that566
dB(t′) = −vS(t′)dt′ − S(0)vh(v)dt′ = −S(t′)dA(t′)− S(0)vh(v)dt′, (C.2)
Using the integration by part formula for stochastic integrals on the product A(t′)S(t′), we have567
−S(t′)dA(t′) = −d(S(t′)A(t′)) +A(t′)dS(t′) (C.3)
since dA(t′)dS(t′) = o(dt′). Consequently,568
dB(t′) = −d(S(t′)A(t′)) +A(t′)dS(t′)− S(0)vh(v)dt′. (C.4)
Integrating (C.4) from 0 to T , and noting A(T−) = A(T ) = 0, gives569
B(T ) = S(0)A(0) +
∫ T
0
A(t′)dS(t′)−∫ T
0
S(0)vh(v)dt′ (C.5)
Note that the last term Bimpact ≡∫ T
0S(0)vh(v)dt′ =
∫ T0S(0)v(S(t′), A(t′), t′)h(v(S(t′), A(t′), t′))dt′ corre-570
sponds to the cost from nonzero trading impact and is stochastic in general. Now we make the assumption571
AS2* The control v(·) is of the form v(·) : (A(t), t) 7→ v, i.e. a static strategy that is independent of S(t).572
With this assumption, the term Bimpact =∫ T
0S(0)v(A(t′), t′)h(v(A(t′), t′))dt′ becomes deterministic. As a573
result, equation (C.5) implies574
V ar
[B(T )
]= V ar
[∫ T
0
A(t′)dS(t′)
]= E
[(∫ T
0
A(t′)dS(t′)
)2]= E
[(∫ T
0
A(t′)σ(S(t′), t′)dW(t′)
)2](C.6)
since the Ito integral has zero expectation. Now we have575
E
[(∫ T
0
A(t′)σ(S(t′), t′) dW(t′)
)2]= E
[∫ T
0
(A(t′)σ(S(t′), t′)
)2dt′]
= E
[(∫ T
0
A(t′)dS(t′)
)2], (C.7)
where the first equality is a result of the Ito isometry.576
References577
Almgren, R. (2003). Optimal execution with nonlinear impact functions and trading enhanced risk. Applied578
Mathematical Finance 10, 1–18.579
Almgren, R. (2009). Optimal trading in a dynamic market. Preprint .580
Almgren, R. and N. Chriss (2000/2001). Optimal execution of portfolio transactions. Journal of Risk 3581
(Winter), 5–39.582
Almgren, R., C. Thum, E. Hauptmann, and H. Li (2004). Direct estimation of market impact. Working583
paper, Citigroup.584
Basak, S. and G. Chabakauri (2010). Dynamic mean-variance asset allocation. Review of Financial Stud-585
ies 23, 2970–3016.586
Bertsimas, D. and A. Lo (1998). Optimal control of execution costs. Journal of Financial Markets 1, 1–50.587
Bjork, T., A. Murgoci, and X. Zhou (2009). Time inconsistent control and mean variance portfolios with588
state dependent risk aversion. PDE and Mathematical Finance III, Stockholm.589
26
Brent, R. (1973). Algorithms for minimization without derivatives. Prentice-Hall, Englewood Cliffs.590
Brugiere, P. (1996). Optimal portfolio and optimal trading in a dynamic continuous time framework. 6’th591
AFIR Colloquium, Nurenberg, Germany.592
Chen, Z. and P. Forsyth (2007). A semi-Lagrangian approach for natural gas storage valuation and optimal593
control. SIAM Journal on Scientific Computing 30, 339–368.594
d’Halluin, Y., P. Forsyth, and G. Labahn (2005). A semi-lagrangian approach for American Asian options595
under jump diffusion. SIAM Journal on Scientfic Computing 27, 315–345.596
Engle, R. and R. Ferstenberg (2007). Execution risk. Journal of Portfolio Management 33 (Winter), 34–44.597
Forsyth, P. A. (2011). A Hamilton Jacobi Bellman approach to optimal trade execution. Applied Numerical598
Mathematics 61, 241–265.599
Gatheral, J. and A. Schied (2010). Optimal trade execution under Geometric Brownian Motion in the600
Almgren and Chriss framework. forthcoming in the International Journal of Theoretical and Applied601
Finance.602
He, H. and H. Mamaysky (2005). Dynamic trading with price impact. Journal of Economic Dynamics and603
Control 29, 891–930.604
Huberman, G. and W. Stanzl (2004). Price manipulation and quasi-arbitrage. Econometrica 72, 1247–1275.605
Lorenz, J. (2008). Optimal trading algorithms: Portfolio transactions, mulitperiod portfolio selection, and606
competitive online search. PhD Thesis, ETH Zurich.607
Lorenz, J. and R. Almgren (2006). Adaptive arrival price. Algorithmic Trading III.608
Lorenz, J. and R. Almgren (2011). Mean variance optimal adaptive execution. to appear, Applied Mathe-609
matical Finance.610
Schied, A., T. Schoeneborn, and M. Tehranchi (2010). Optimal basket liquidation for CARA investors is611
deterministic. Applied Mathematical Finance 17, 471–489.612
Schied, A. and T. Schoneborn (2008). Risk aversion and the dynamics of optimal liquidation strategies in613
illiquid strategies in illiquid markets. Working paper, Technical University of Berlin.614
Tse, S. T., P. A. Forsyth, J. S. Kennedy, and H. Windcliff (2011). Comparison between the mean vari-615
ance optimal and the mean quadratic variation optimal trading strategies. Working paper, University of616
Waterloo.617
Vath, V. L., M. Mnif, and H. Pham (2007). A model of optimal portfolio selection under liquidity risk and618
price impact. Finance and Stochastics 11, 51–90.619
Wang, J. and P. Forsyth (2008). Maximal use of central differencing for Hamilton-Jacobi-Bellman PDEs in620
finance. SIAM Journal on Numerical Analysis 46, 1580–1601.621
Wang, J. and P. Forsyth (2010). Numerical solution of the Hamilton Jacobi Bellman formulation for contin-622
uous time mean variance asset allocation. Journal of Economic Dynamics and Control 34, 207–230.623
Wang, J. and P. Forsyth (2011a). Comparison of mean variance like strategies for optimal asset allocation624
problems. To appear, International Journal of Theoretical and Applied Finance.625
Wang, J. and P. Forsyth (2011b). Continuous time mean variance asset allocation: a time consistent strategy.626
European Journal of Operational Research 209, 184–201.627
27